IBM Releases Granite Switch 4.1 30B — a Modular LLM with Adapters and Speech Recognition

The model ibm-granite/granite-switch-4.1-30b with the text-generation task appeared on HuggingFace on October 14, 2026. It is a modular LLM that combines a base model, two adapters and an audio encoder in a single endpoint: one hosting can answer with text, make decisions and recognize English speech. The adapters are enabled on demand via control tokens in the chat template.

What it means

The model contains 30 billion parameters; all weights in the repository take up 55.9 GB in safetensors format. The largest file is 4.3 GB. License Apache 2.0, launch via HuggingFace Transformers or vLLM. There are no quantized builds on the platform yet, so running locally will require a GPU with enough VRAM for the full weight of the model. It is based on granite-4.1-30b with a 128K context. Two adapters are connected on top of it: decision for general decisions and guardian-core from granitelib-guardian-r1.0. The audio part is granite-speech-5.0-470m-turboctc with 470 million parameters for English speech recognition. The same line includes the lighter granite-switch-4.1-8b based on 8B. You can follow news about tools for running on your own hardware in the CPU3D news feed.